My first attempt at figuring out what someone wants to watch was a rule. Fewer than five words in the search box, they probably want a specific movie. More than five, they're probably describing a vibe.

(I know. Stay with me.)

Two jobs, one search box

Tasty Mangoes started as a way to manage and rotate streaming subscriptions. Pretty quickly I realized that the bigger problem is the step before that: figuring out what you actually want to watch. And that turned out to be two completely different problems wearing the same outfit.

Sometimes you know exactly what you're looking for. You type a title, you find it, done. That's a search.

Other times you have no idea. You want something funny, but not dumb, and short enough that you won't fall asleep halfway through. That's not a search. That's a feeling you're trying to describe.

One box, two jobs. So I tried to make the box smart enough to tell which one you meant, and the five-word rule was my best guess.

What happened

Tester momentone specific thing a tester typed or did that broke the rule, or that you didn't expect. Even one sentence works.

As you can imagine, counting words only gets you so far. People don't type like a rule says they should.

So I stopped guessing and split it in two. One tool for finding a specific movie. Another for semantic search, where you describe what you want in plain English and each result explains why it showed up. The user decides which one they need.

That worked. But it also made something else obvious.

The thing underneath

People don't think about picking a movie in neat categories, and they usually don't know all their criteria when they start. Maybe you begin with "something funny," then see a result and realize you also want it to be short. Then you change your mind about funny entirely.

A search like that isn't a form you fill out. It evolves. And evolving things don't mix well with rigid rules.

Picking a movie, it turns out, is a lot less like searching and a lot more like talking to a friend.

Where it went

Right around then, OpenAI's real-time voice technology got really good, and the timing was too good to ignore. That became Mango Live: you talk to Mango while you look for something to watch, change your mind halfway through, and movie cards pop up on screen so you can look at the poster or tap a button whenever that's easier than saying it out loud.

It's still rough in places. Voice is also expensive at scale, so instead of rushing into a wider beta, I slowed down on purpose to understand the economics first. Three steps forward, one step back.

What I took from it

Your rolewho came up with the five-word rule, who built the two tools, what you personally designed, tested, or decided versus your dad, contractors, or AI tools.

The pattern I keep running into is that I start with a narrow, practical function (rotate your subscriptions, search for a title) and find a human problem underneath it (what do I even want?). The more I build, the more I think the real design work is in that second layer.

What i want to understand nextone or two sentences. For example, what a good "friend" remembers about you, and what it should forget.